注意力机制量化剪枝优化方法

面向基于注意力机制模型的巨大计算和访存开销问题,研究量化和剪枝协同优化的模型压缩技术,提出针对注意力机制中查询、键、值、概率共四个激活值矩阵的对称线性定点量化方法。同时,提出概率矩阵剪枝方法和渐进式剪枝策略,有效降低剪枝精度损失。在不同数据集上的实验结果表明,针对典型基于注意力机制模型BERT,在较低或者没有精度损失的情况下该优化方法可达到4位或8位定点量化、0.93～0.98的稀疏度,大幅度降低模型计算量,为加速量化稀疏模型的推理奠定良好的基础。     

Effect of nano powder (Al2O3-CaO) addition on the mechanical properties of the polymer blend matrix composite

This work describes the preparation and study of the properties of composite nanoparticles prepared by the sol-gel method which consists of two materials (Al₂O₃-CaO), and study the effect of these nanoparticles on the mechanical behavior of a polymer blend (EP 4% + 96% UPE). The powder was evaluated by X-ray diffraction analysis, scanning electron microscopy analysis (SEM), particle size analysis, and energy dispersive X-ray analysis (EDX). The mechanical behavior of the composite material was assessed by tensile test, bending test and hardness test. The evaluation results of the composite nanoparticles showed good distribution of the chemical composition between aluminum oxide and calcium oxide, smoothness in particles' size at calcination in high and low temperatures, formation of different shapes of nanoparticles and different (kappa and gamma) phases of the Al₂O₃ particles. The results of mechanical behavior tests showed marked improvement in the mechanical properties of the resulted composite material, especially at 1.5%, compared with polymer blend material without nano powder addition. The tensile properties improved about (24 and 14.9) % and bending resistance about (23.5 and 16.8) % and hardness by (25 and 22) % when adding particles of size (63.8 and 68.6) respectively. Therefore, this reflects the efficiency of the proposed method to manufacture the nanocomposite powder and the possibility of using this powder as a strengthening material for the composite materials and using these composite materials in bio applications, especially in the fabrication of artificial limbs.     

干扰抑制类生成式对抗网络

为进一步改善超低频频段的通信质量,在传统改进广义旁瓣抵消算法的基础上,提出新的超低频干扰抑制算法——生成式旁瓣抵消算法。该算法将人工智能研究热点之一的生成式对抗网络模型引入广义旁瓣抵消算法中,通过优化设计生成模型的网络结构及相关超参数,有效地解决了原算法存在的期望信号残留问题,为旁瓣抵消通道中的后级滤波算法提供了与主通道相关性更强的干扰参考信息,从而提高了算法对主通道干扰估计的准确性。为了验证优化后生成模型的有效性以及所提算法对不同类别干扰的抑制能力,在实验室环境下搭建实验平台,设计了多组对照实验。实验结果表明:优化后的生成模型具有较好的生成能力、较好的鲁棒性以及相对较低的运算复杂度;相比于传统改进的广义旁瓣抵消算法,所提算法进一步提高了信号带宽内的信干噪比。     

模糊时滞系统控制器依赖于状态时滞的H∞控制

针对一类非线性时滞系统,研究了该系统基于T—S模糊模型的H∞控制器设计问题。采用线性矩阵不等式LMI的方法,设计一个依赖于状态时滞的模糊控制器,得到了系统存在模糊控制器的充分条件,此充分条件等价于一类线性矩阵不等式的可解性,最后通过仿真说明了控制器的有效性。     

非线性时滞系统模糊控制的一种新方法

针对一类状态变量具有时滞的非线性系统,研究了其模糊控制器的设计问题。采用线性矩阵不等式（LMI）的方法,设计了与状态变量导数有关的模糊控制器,得到了系统存在模糊控制器的充分条件,并利用线性矩阵不等式的解给出具体的形式,最后通过仿真说明了控制器的有效性。     

基于最大故障特征信息熵的测试点优化设计

为解决故障诊断和可测试性设计中普遍存在的测试点优化问题,以相关性矩阵模型为基础,建立了最大故障特征信息熵的优化准则,给出了相应的测试点优化设计方法。应用实例表明：该方法是可行的、正确的,能为武器装备的故障诊断和可测试性设计提供有力的支持。     

饱和系统的线性最优容错不变集

针对饱和系统的容错控制问题,提出了一种D稳定约束下系统吸引域的估计方法。分别以不变集和参考集的容量为优化目标,给出了两个优化容错不变集的充分条件。为解决不变集优化中出现的双线性矩阵不等式(BMI)问题,通过构造包含线性矩阵不等式(LMI)的目标函数,将它转化为非线性规划的形式进行求解。通过一个仿真实例对该方法进行验证。     

感应式磁性天线设计及干扰抑制算法研究

当期望信号和干扰同方向时,为了有效改善超低频频段的通信质量,提出了一种基于模拟电路预处理和改进广义旁瓣抵消的干扰抑制算法,设计了磁性天线、低噪声前置放大电路,制作了灵敏度较高的磁传感器,有效地抑制了工频及其谐波干扰。鉴于超低频频段的信号十分微弱,在广义旁瓣抵消算法的基础上做了几点改进,为主通道提供较多的参考信息,从而提高了算法的性能,有效地解决了传统算法失效的问题。为了验证所提算法的有效性,在实验室环境下搭建实验平台,设计了多组对照实验,实验结果表明:无论期望信号与干扰是否同方向,改进后的广义旁瓣抵消算法相比原来的算法,在信噪比的提升和噪声底限的降低等方面均有较大程度的改善。     

Fabrication of AA7005/TiB2-B4C surface composite by friction stir processing: Evaluation of ballistic behaviour

The present work aims to enhance the ballistic resistance of AA7005 alloy by incorporating the TiB₂ and B₄C ceramic reinforcement particles. Surface composites with different weight fractions of TiB₂ and B₄C particles were processed by friction stir processing. Micro-hardness and depth of penetration tests were carried out to evaluate the ballistic properties of the surface composites. The surface hardness of the composite was found to be nearly 70 HV higher than base alloy. The depth of penetration of the steel projectile was 20–26 mm in the composites as compared to 37 mm in the base alloy. Ballistic mass efficiency factor of the surface composite was found to be 1.6 times higher than base alloy. This is mainly attributed to the dispersion strengthening from the reinforcement particles.     

基于面积比矩阵的加权特征向量点匹配算法

提出一种在仿射变换下点集关系描述算子--面积比矩阵,证明了其在仿射变换群下的不变性质;并由该性质推导得出特征向量(EA)匹配算法抗噪性能差的原因,在此基础上给出了基于面积比矩阵的加权特征向量算法.算法构造不同图像的面积比矩阵,对矩阵进行分解得到其特征值和特征向量,通过特征值进行加权获得图像中点的特征向量,比较图像点的特征向量获取匹配关系.因面积比矩阵的仿射不变性质,算法能实现仿射变换下点集精确匹配;采用的加权特征向量法改进了EA匹配法,具有更好的抗噪性能.实验表明算法切实可行.